Jason Ansel
Jason Ansel
I think this is unlikely to be fixed given the [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/) direction, though PRs are welcome to fix this issue.
That is correct. Maybe we would need a pattern for the backwards as well?
> I also found that `ExternKernelCaller` has a `has_out_variant` attribute It will be used when this one lands https://github.com/pytorch/pytorch/pull/95556
> > Those timeouts in CI are a bit worrying. I wonder if this could be going into an infinite loop to cause them. > > I was also concerned...
When `suppress_errors=True` (the default for the PYTORCH_TEST_WITH_DYNAMO=1 tests) then `AssertionError` results in falling back to eager mode and the test passing. If we hit the assertion you could print out...
Errors are here: https://hud.pytorch.org/pr/97020 Looks like infinite recursion on one from L3587.
> I have a question, can a decomposition be added for ops from `prims` namespace? @jansel @davidberard98 Yes, you can decompose prims -- but those should go in `_inductor/decomposition.py`
> > Errors are here: https://hud.pytorch.org/pr/97020 > > Looks like infinite recursion on one from L3587. > > Yeah, I don't understand why that's happening. Is there a command to...
@pytorchbot merge
> > This is my other PR: #93919, waiting for approval for a long time. Can you please review that as well? > > @jansel Also, could you please take...